Neural Cinema: explore the path of a signal
Neural Cinema is an interactive Three.js visual metaphor for neural-network-style signal flow. It generates animated layouts for a demonstration; it does not train a model, inspect a live model, or show measured token probabilities.
A short demonstration you can repeat
- Open the demo and select MLP. Follow the layered path from the request and intent through the policy bottleneck to the tool, risk and action branches.
- Pause the animation and identify a narrow layer and a fan-out layer. Resume to follow the moving pulses.
- Keep the same preset and move Density. Observe the change in visible connections, then reseed to produce a different generated layout.
- Select Vision, then Language. Compare local feature mixing with longer skip connections. Discuss what the arrangement suggests, without treating it as evidence about an actual AI model.
For a classroom or recorded explanation, keep the preset and density visible. Reseeding changes the layout, so screenshots from different seeds are not a controlled measurement of network performance.
What do the three presets represent?
| Preset | Visual idea to discuss |
|---|---|
| Agent Decision MLP | A policy bottleneck followed by tool, risk and action gates. |
| Vision Feature Mixer | Local receptive-field-style mixing from pixels toward a class. |
| Language Router | Memory and attention-style skip links toward token selection. |
These are authored topology presets, not implementations of a production policy, vision model or language model.
What do Density, Signal and Activation mean here?
Density changes the generated connection layout. It does not change an actual model's training data, accuracy or compute budget. Reseed generates a new arrangement; pause freezes the animation clock.
The layer readout displays an activation label and a signal value. Node activation and the readouts come from the demo's procedural animation, not recorded training or inference data. Use them to follow motion through the illustration, not to diagnose a model.
FAQ
Should I use this or TensorFlow Playground?
Use Neural Cinema when you want a visual explanation of layered paths, bottlenecks and skip links. TensorFlow Playground exposes a dataset, learning rate and training/test loss for experimenting with a small neural network. Those training controls are not features of Neural Cinema.
Why is the canvas blank or showing a dependency error?
The demo requires WebGL and pinned CDN builds of Three.js and Lucide. Confirm that the browser supports WebGL and that the CDN requests can load. A blocked script, offline first visit or browser graphics restriction may stop the animation; the page shows a dependency error when Three.js is unavailable.
Does running locally remove the need for an internet connection?
No. Serving this directory locally still uses the CDN dependencies. The README explains the static-server setup and existing checks.
Where is the implementation and how can I report a problem?
app.js defines the presets, generated connections and animation. Include the browser, preset and visible error in an issue; do not attach credentials. The project has an MIT license.